On January 27, 2025, an unknown app took the top spot on the US App Store. Named DeepSeek and hailing from Hangzhou, it was free. In its wake, the Nasdaq tumbled 3%, Nvidia shed 17% in a single session, wiping out $589bn in market cap, the biggest wipeout of shareholder value ever recorded in a single day.
Fifteen months later, history repeats itself. Last Friday, DeepSeek launched V4 in preview. Open source, boasting 1.6 trillion parameters, and trained on Chinese Huawei Ascend 950 and Cambricon chips. Following the news, SMIC rose 10% in Hong Kong, while Hua Hong jumped 15%. However, this time, the Nasdaq didn't flinch. Between coffees in the newsroom, we have been wondering for some time: where exactly does Chinese AI stand?
The silence has just ended
In a statement DeepSeek claims that V4-Pro "rivals the world's best proprietary models" and outperforms all open-source models in agentic coding. It remains surpassed only by Gemini 3.1 Pro in general knowledge. According to CNN, V4 was trained on Chinese Huawei Ascend 950 and Cambricon chips, not Nvidia. We will return to this.
It remains to be seen how the model performs under real-world conditions: the version is still in preview. DeepSeek itself acknowledges a 3-to-6 month lag behind GPT-5.4 and Gemini 3.1 Pro. However, the signal is clear: China has just proven it can deliver cutting-edge performance via open source.
And Alibaba is not trailing behind. According to the official technical sheet published on Hugging Face, its Qwen 3.5-9B, released on March 2, scored 81.7 on GPQA Diamond, a PhD-level reasoning test. OpenAI's gpt-oss-120b, which carries 13x more total parameters (the internal settings determining a model's size), peaked at 80.1. Behind these advances lies a paradox.
The dollar time lag
According to the Stanford AI Index 2026, the US injected $285.9bn in private AI investment in 2025, compared to $12.4bn in China, i.e. a 23-fold difference. And the gap is set to widen: the four American hyperscalers alone plan nearly $700bn in AI infrastructure investment for 2026. US soil already hosts 5,427 data centers, ten times more than any other country.
Yet, the more the Americans spend, the more the gap narrows. According to Artificial Analysis, two Chinese models (GLM-5.1 and DeepSeek V3.2) now feature in the global top 10. DeepSeek V3.2 is 33 times cheaper to use than Claude Opus 4.7 (as we detailed in a previous article). V4-Pro, released Friday, drops to $3.48 per million output tokens, compared to $25 at Anthropic, according to Fortune. However, while the focus remains on benchmarks and billions raised, the real game is being played elsewhere.
When Beijing locks everything down
This is where the landscape becomes complicated. Earlier this week, Beijing officially blocked the $2bn acquisition of Manus—a Chinese agentic AI gem—by Meta, according to Bloomberg. This is the culmination of an investigation launched in January, after the two co-founders were banned from leaving the country.
The message to Chinese entrepreneurs is crystal clear: you can grow, but you cannot escape.
And the lockdown does not stop there. On April 22, US Commerce Secretary Howard Lutnick admitted to the Senate that Nvidia has yet to sell a single H200 chip to China, despite Trump giving the green light in January. The reason? Beijing itself is refusing authorization, "to keep its investments focused on its own domestic industry," according to Reuters.
The reversal is striking: it is no longer Washington blocking; it is now Beijing closing the door. And that changes everything.
The other signal is the local value chain that China is deploying at an accelerated pace. Take BYD, a global leader in electric vehicles. Its autonomous driving system, standardized across all models including the $9,500 Seagull, relies on DeepSeek R1 AI.
Chinese car, Chinese AI, Chinese consumers. Only one link remains missing: the chips, currently a mix of Nvidia and Horizon Robotics.
So, where is this Chinese AI?
Everywhere. In the Ascend chips that Huawei produces with its Chinese partner SMIC, sometimes using TSMC dies obtained through indirect channels according to the RAND Corporation. In the startups that Beijing now prevents from leaving or being sold. In the models that rival the best American offerings for 33 times less.
With one caveat, however: according to an FT investigation reported by CNBC in July 2025, over $1bn worth of export-banned Nvidia chips continued to enter China via parallel channels within a 3-month period.
Is Beijing blocking American chips out of pure patriotism, or simply to force its industry forward while relying on the black market to avoid falling behind?
Regardless, America has not lost the game yet: three American models still lead the Artificial Analysis rankings (Claude Opus 4.7, Gemini 3.1 Pro, and GPT-5.4). But the era of comfort is over.
China is not chasing Silicon Valley. It is building its own ecosystem with its own local stars. And while the Americans are closing their models (Meta launched Muse Spark, its first proprietary model, in early April after the failure of Llama), China is releasing them as open source.
DeepSeek, Qwen, GLM-5.1—all free, all downloadable. Chinese low-cost is not just about price. It is also a strategy of influence.
https://www.marketscreener.com/news/where-has-chinese-ai-gone-ce7f59d2dd8bfe23
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.